Add SSE keepalives during long image generation calls

[adambalogh]

Summary

Image generation requests to providers take 60-120 seconds with no bytes on the wire. Idle-timeout proxies between the enclave and browser reset such silent streams at ~60s, surfacing as ERR_HTTP2_PROTOCOL_ERROR on the client. This change emits SSE comment frames ("keepalives") at regular intervals while generation runs to keep bytes flowing and prevent proxy timeouts.

Key Changes

• New run_with_sse_keepalives() generator (image_generation.py): Runs a callable on a daemon worker thread, yielding SSE comment frames (: keepalive\n\n) at configurable intervals while the work executes. Returns the callable's result or re-raises any exception. Designed for use with yield from in streaming response generators.

• Updated image generation timeout (image_generation.py): Added _IMAGE_GENERATION_TIMEOUT (300s read, 15s connect) to both /images/generations POST calls (form-based edits and JSON generations). Image generation is far slower than chat and needs its own generous timeout instead of the shared client default.

• Integrated keepalives into streaming response (image_generation.py): Modified create_image_generation_streaming_response() to wrap _run_image_generation() with run_with_sse_keepalives(), ensuring keepalive frames flow while the provider call executes. Final signed frame and error semantics unchanged.

• Extended keepalives to Gemini image generation (chat_controller.py): Applied the same keepalive pattern to the Gemini image generation path in generate() where model.invoke() is called for image-output models, protecting that long-running call as well.

• Comprehensive test coverage (test/test_image_generation.py): Added TestStreamingKeepalives class with four test cases:

  • test_keepalives_flow_while_generation_runs: Verifies keepalives are emitted during slow generation and final signed frame is intact
  • test_generation_error_still_yields_error_frame: Confirms error frames are properly emitted even with keepalives
  • test_fast_generation_emits_no_keepalives: Ensures no unnecessary keepalives for quick operations
  • test_run_with_sse_keepalives_returns_value_and_raises: Unit tests the keepalive generator's return value and exception propagation

Implementation Details

• Keepalive interval is configurable via _KEEPALIVE_INTERVAL_SECONDS (default 10s, patched to 0.01s in tests for speed)
• Worker thread is daemon-threaded; if the client disconnects mid-generation, the generator closes but the provider call continues on the daemon
• OHTTP encrypter downstream holds each chunk until the next arrives (one-chunk look-ahead), so wire sees bytes at most every 2x the interval—still far under the ~60s idle timeouts
• Comment frames are SSE-compliant and ignored by parsers (the app's stream parser only reads data: lines)

https://claude.ai/code/session_017bwXiKz9YQsuhTyRcwEnKz